The GraphNet Zoo: An All-in-One Graph Based Deep Semi-Supervised Framework for Medical Image Classification

We consider the problem of classifying a medical image dataset when we have a\nlimited amounts of labels. This is very common yet challenging setting as\nlabelled data is expensive, time consuming to collect and may require expert\nknowledge. The current classification go-to of deep supervised learning is\nunable to cope with such a problem setup. However, using semi-supervised\nlearning, one can produce accurate classifications using a significantly\nreduced amount of labelled data. Therefore, semi-supervised learning is\nperfectly suited for medical image classification. However, there has almost\nbeen no uptake of semi-supervised methods in the medical domain. In this work,\nwe propose an all-in-one framework for deep semi-supervised classification\nfocusing on graph based approaches, which up to our knowledge it is the first\ntime that an approach with minimal labels has been shown to such an\nunprecedented scale with medical data. We introduce the concept of hybrid\nmodels by defining a classifier as a combination between an energy-based model\nand a deep net. Our energy functional is built on the Dirichlet energy based on\nthe graph p-Laplacian. Our framework includes energies based on the $\\ell_1$\nand $\\ell_2$ norms. We then connected this energy model to a deep net to\ngenerate a much richer feature space to construct a stronger graph. Our\nframework can be set to be adapted to any complex dataset. We demonstrate,\nthrough extensive numerical comparisons, that our approach readily compete with\nfully-supervised state-of-the-art techniques for the applications of Malaria\nCells, Mammograms and Chest X-ray classification whilst using only 20% of\nlabels.\n